The Agent Skills Directory

[UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The skill uses npx to execute the vendor-specific CLI tool (@tiangong-ai/cli). The package is version-pinned and belongs to the authoring organization, which represents standard functionality for this platform.
[PRIVILEGE_ESCALATION]: The skill integrates with system-level isolation tools such as sandbox-exec on macOS and bwrap (Bubblewrap) on Linux. These tools are employed as defensive measures to sandbox research agents and enforce the principle of least privilege during execution.
[DATA_EXPOSURE_AND_EXFILTRATION]: The skill handles sensitive API keys for research services but implements several protective measures. Its documentation recommends strict file permissions (chmod 600) for credential files and uses a brokered network approach that prevents sensitive tokens from being logged or included in research outputs.
[INDIRECT_PROMPT_INJECTION]: As a research tool that ingests data from external sources (such as via brokered-network in SKILL.md), there is a theoretical injection surface. The skill addresses this through architectural safeguards: it uses 'bounded views' to limit the amount of external content exposed to the LLM, employs content-addressed receipts for integrity, and requires a multi-stage review process that verifies findings against hash-bound raw evidence.

tiangong-auto-research